Customer-controlled data plane
Customer data and pipeline processing remain in the customer-managed cloud and run on the customer's Databricks or Snowflake compute.
MCP security
Technical controls for governed agent access
DataForge gives AI agents a controlled path into the data platform without creating a separate AI bypass. Identity, authorization, validation, and destructive-action controls remain enforced by the platform.
These are technical controls, not third-party security certifications. DataForge will provide current architecture details and control evidence during a security review.
The control model
The MCP server exposes DataForge capabilities through the same governed platform used by human operators. The controls below describe how the current product constrains access and changes.
Authenticated agent access
OAuth 2.1 with PKCE supports interactive MCP clients. Agents operate as an authenticated user rather than through a shared, privileged AI identity.
Authorization on every call
Project roles and database-enforced permissions are checked when tools run. An AI client does not receive a separate authorization path.
Validation before persistence
Agent and human changes pass the same save-time schema and expression validation. Invalid writes are rejected before they become platform objects.
Server-enforced confirmation
Destructive operations require explicit confirmation enforced by the server, not merely requested through an agent prompt.
Constrained data access
Read-only behavior is the default where applicable. SQL access to customer data is separately controlled and off by default until the customer opts in.
Operational evidence
Talos and MCP activity use the same tool registry and telemetry model. Outcomes, latency, release version, and argument or result shape can be recorded without logging customer values.
No AI bypass
External MCP clients, Talos, and human operators act through the platform's established controls. AI does not receive reduced validation or a forgiving write path.
For security and architecture teams
Take the overview into your review
The MCP Security and Control Overview summarizes the trust boundaries, access controls, validation path, data posture, and questions your team should verify for its deployment.
Get the security overviewReview your architecture with us
Bring your identity, network, Databricks or Snowflake, and AI-client requirements. We will walk through the relevant trust boundaries and controls with your team.
Request an architecture review